US11906441B2 - Inspection apparatus, control method, and program - Google Patents

Abstract

An inspection apparatus (100) detects an inspection object (90) from first image data (10) in which the inspection object (90) is included. The inspection apparatus (100) generates second image data (20) by performing a geometric transform on the first image data (10) in such a way that a view of the detected inspection object (90) becomes a view satisfying a predetermined reference. In an inference phase, the inspection apparatus (100) detects, by using an identification model for detecting an abnormality of the inspection object (90), an abnormality of the inspection object (90) included in the second image data (20). Further, in a learning phase, the inspection apparatus (100) learns, by using the second image data (20), an identification model for detecting an abnormality of the inspection object (90).

Description

This application is a National Stage Entry of PCT/JP2019/022037 filed on Jun. 3, 2019, the contents of all of which are incorporated herein by reference, in their entirety.

TECHNICAL FIELD

The present invention relates to an inspection of a structure and the like.

BACKGROUND ART

As working population decreases, there is an increasing demand for assisting or automating, by means of an image recognition technique, determination as normal or abnormal made by a skillful maintenance worker who carries out an inspection, a test, or the like. In image recognition, it is necessary to improve detection accuracy by using a large volume of learning images (learning data) collected on site. However, when an inspection object (for example, a power transmission line, a bolt, an insulator, a railroad rail, a sleeper, an overhead wire, and the like) has various patterns of views (for example, a position, an inclination, a size, and the like of the inspection object in an image), that is, various patterns of appearances in an image, there is a high possibility that a model for detecting an abnormal part included in the inspection object becomes an unlearned state (underfitting) where learning is insufficient, or a state of ending up with a local solution rather than an optimum solution. Consequently, sufficient detection accuracy cannot be acquired.

PTL 1 discloses a patrol inspection system that determines an arc mark generated by a stroke of lightning on an overhead ground wire or a power transmission line as abnormal and stably detects the arc mark by means of image processing. PTL 2 discloses a technique for normalizing a contrast of an image to be tested and inspecting deterioration of a concrete surface by using the normalized image.

RELATED DOCUMENTPatent Document

[PTL 1] Japanese Patent Application Publication No. 2018-074757

[PTL 2] Japanese Patent Application Publication No. 2000-028540

SUMMARY OF THE INVENTIONTechnical Problem

In general, an inspection object has mostly the same shape. However, a view (a shape, an angle, a size, or the like) of the inspection object largely varies depending on a structure, a capturing method, and the like. In other words, there are various patterns of appearances of the inspection object in an image. Accordingly, there are also various patterns of views of an abnormal part to be detected. Thus, in order to learn a model for achieving high detection accuracy by using a deep-learning technique and the like, it is required that a large volume of learning data covering all the patterns of views of the abnormal part are collected.

However, with increase in patterns to be learned, there is a high possibility that a model for detecting an abnormal part becomes an unlearned state or a state of ending up with a local solution. Consequently, sufficient detection accuracy cannot be acquired. PTLs 1 and 2 do not mention the above-described problem.

The present invention has been made in view of the above-described problem, and one of objects of the present invention is to provide a technique for improving accuracy in detection of an abnormal part.

Solution to Problem

A first inspection apparatus according to the present invention includes: 1) a transformation unit that detects an inspection object from first image data in which the inspection object is included, and generates second image data by performing a geometric transform on the first image data in such a way that a view of the detected inspection object becomes a view satisfying a predetermined reference; and 2) a detection unit that detects, by using an identification model for detecting an abnormality of the inspection object, an abnormality of the inspection object included in the second image data.

The identification model is learned by using image data in which an inspection object having a view satisfying a predetermined reference is included.

A second inspection apparatus according to the present invention includes: 1) a transformation unit that detects an inspection object from first image data in which the inspection object is included, and generates second image data by performing a geometric transform on the first image data in such a way that a view of the detected inspection object becomes a view satisfying a predetermined reference; and 2) a learning unit that learns, by using the second image data, an identification model for detecting an abnormality of the inspection object.

A first control method according to the present invention is executed by a computer. The control method includes: 1) a transformation step of detecting an inspection object from first image data in which the inspection object is included, and generating second image data by performing a geometric transform on the first image data in such a way that a view of the detected inspection object becomes a view satisfying a predetermined reference; and 2) a detection step of detecting, by using an identification model for detecting an abnormality of the inspection object, an abnormality of the inspection object included in the second image data.

The identification model is learned by using image data in which an inspection object having a view satisfying a predetermined reference is included.

A second control method according to the present invention is executed by a computer. The control method includes: 1) a transformation step of detecting an inspection object from first image data in which the inspection object is included, and generating second image data by performing a geometric transform on the first image data in such a way that a view of the detected inspection object becomes a view satisfying a predetermined reference; and 2) a learning step of learning, by using the second image data, an identification model for detecting an abnormality of the inspection object.

A program according to the present invention causes a computer to execute each of steps in the first control method or the second control method according to the present invention.

Advantageous Effects of Invention

According to the present invention, a technique for improving accuracy in detection of an abnormal part is provided.

BRIEF DESCRIPTION OF THE DRAWINGS

The above-described object and other objects, features, and advantageous effects become more apparent from the preferred example embodiments described below and the following accompanying drawings.

FIG.1 is a diagram representing an overview of an operation of an inspection apparatus according to an example embodiment 1.

FIG.2 is a diagram illustrating first image data.

FIG.3 is a diagram illustrating the first image data.

FIG.4 is a diagram illustrating a configuration of the inspection apparatus according to the example embodiment 1.

FIG.5 is a diagram illustrating a computer for achieving the inspection apparatus.

FIG.6 is a flowchart illustrating a flow of processing executed in an inference phase by the inspection apparatus according to the example embodiment 1.

FIG.7 is a flowchart illustrating a flow of processing executed in a learning phase by the inspection apparatus according to the example embodiment 1.

FIG.8 illustrates one example of teaching data.

FIG.9 illustrates one example of different teaching data.

FIG.10 illustrates one example of a result of detecting a line on the first image data.

FIG.11 illustrates one example of second image data according to the present example embodiment.

FIG.12 is a block diagram illustrating a function configuration of the inspection apparatus including an output unit.

FIG.13 is a diagram illustrating one example of output information.

FIG.14 is a diagram illustrating the second image data on which a display representing an abnormal part is superimposed.

FIG.15 is one example of a degree of overlapping of patch images illustrated by a heat map.

FIG.16 is a block diagram illustrating a function configuration of the inspection apparatus including a checking unit.

FIG.17 is a diagram illustrating a screen output by the checking unit.

DESCRIPTION OF EMBODIMENTS

Hereinafter, example embodiments of the present invention will be described by using the drawings. Note that, a similar component is assigned with a similar reference sign throughout all the drawings, and description therefor will be omitted as appropriate. Further, in each block diagram, each block represents not a configuration on a hardware basis but a configuration on a function basis, except as particularly described.

Example Embodiment 1

FIG.1 is a diagram representing an overview of an operation of aninspection apparatus100 according to an example embodiment 1. Note that,FIG.1 is illustrative for ease of understanding theinspection apparatus100, and a function of theinspection apparatus100 is not limited to the illustration inFIG.1.

Theinspection apparatus100 detects, by usingfirst image data10 acquired by capturing an image of aninspection object90, an abnormality relating to theinspection object90. Thefirst image data10 are image data acquired by capturing an image of theinspection object90.

FIGS.2 and3 are diagrams illustrating thefirst image data10. Both of thefirst image data10 inFIG.2 and thefirst image data10 inFIG.3 include a power transmission line11, an arc mark12, and a snow-resistant ring13. Herein, the arc mark12 is an abnormal part. A frame14 represents an area (hereinafter, an analysis area) as a unit for processing in an image analysis.

Herein, a view of theinspection object90 on thefirst image data10 may vary depending on a position, a pose, and the like of a camera that captures an image of theinspection object90. The view of theinspection object90 on thefirst image data10 is a position, an angle, a size, and the like of theinspection object90 on thefirst image data10. For example, when thefirst image data10 inFIG.2 are compared with thefirst image data10 inFIG.3, shapes of the power transmission lines are mutually the same, whereas inclinations and sizes of the power transmission lines in the images are mutually different. Thus, a view of the abnormal part may also differ for each image.

In this regard, theinspection apparatus100 generates, by performing a geometric transform on thefirst image data10, image data (second image data20) in which a view of theinspection object90 satisfies a predetermined reference. In this way, variation in views of theinspection object90 included in image data to be analyzed is reduced.

For detection of an abnormality of theinspection object90, an identification model is used. The identification model is a model learned in such a way that an abnormality of theinspection object90 included in image data is detected, and is stored in a storage apparatus accessible from theinspection apparatus100.

In a learning phase for learning the identification model, theinspection apparatus100 learns the identification model by using thesecond image data20. With this, the identification model can detect an abnormality of theinspection object90 from image data in which theinspection object90 having a view satisfying a predetermined reference is included.

Further, in an inference phase (application phase) for detecting an abnormality of theinspection object90 by use of the learned identification model, theinspection apparatus100 detects an abnormality of theinspection object90 by using thesecond image data20 and the identification model. As described above, the identification model is learned in such a way that an abnormality of theinspection object90 can be detected from image data in which theinspection object90 having a view satisfying a predetermined reference is included. Thus, theinspection apparatus100 can detect an abnormality of theinspection object90 by using thesecond image data20 and the identification model.

One Example of Advantageous Effect

It is assumed that an identification model is learned by using thefirst image data10 as is, without performing the geometric transform described above. In this case, variation in views of theinspection object90 is large for each piece of thefirst image data10, and, in order to learn the identification model in such a way as to cover such a large variety of views of theinspection object90, a large volume of learning data (the first image data10) is required. Thus, there arises a problem that long time is required for learning or large labor is required for preparing a large volume of learning data.

In this regard, theinspection apparatus100 generates, by performing a geometric transform on thefirst image data10, thesecond image data20 in which a view of theinspection object90 satisfies a predetermined reference, and learns the identification model by using thesecond image data20. Therefore, the identification model is learned with learning data having small variation in views of theinspection object90, and thus, the number of pieces of learning data required for learning the identification model can be reduced. This has an advantageous effect that time required for learning the identification model is shortened and an advantageous effect that the number of pieces of learning data required for learning the identification model can be reduced.

Further, the identification model learned with learning data having small variation can be said to be high in accuracy for identification, in comparison with the identification model learned with learning data having large variation. In this regard, theinspection apparatus100 generates thesecond image data20 from thefirst image data10, and detects an abnormality of theinspection object90 by using thesecond image data20 and the identification model. Thus, an abnormality of theinspection object90 can be detected with high accuracy, in comparison with a case in which an abnormality of theinspection object90 is detected directly from thefirst image data10.

Hereinafter, theinspection apparatus100 according to the present example embodiment will be described in further detail.

Example of Function Configuration ofInspection Apparatus100

FIG.4 is a diagram illustrating a configuration of theinspection apparatus100 according to the example embodiment 1. Theinspection apparatus100 includes atransformation unit120, alearning unit130, and adetection unit150. Thetransformation unit120 generates thesecond image data20 by applying a geometric transform to thefirst image data10 in such a way that a view of theinspection object90 satisfies a predetermined reference. Thedetection unit150 detects, by using thesecond image data20 and an identification model, an abnormality of theinspection object90 included in thesecond image data20. Thelearning unit130 learns an identification model by using thesecond image data20.

Herein, theinspection apparatus100 used in the learning phase and theinspection apparatus100 used in the inference phase may be separate apparatuses. In this case, theinspection apparatus100 used in the learning phase includes thetransformation unit120 and thelearning unit130. On the other hand, theinspection apparatus100 used in the inference phase includes thetransformation unit120 and thedetection unit150.

Hardware Configuration ofInspection Apparatus100

Each of function configuration units of theinspection apparatus100 may be achieved by hardware (example: a hard-wired electronic circuit, or the like) achieving each of the function configuration units, or may be achieved by a combination of hardware and software (example: a combination of an electronic circuit and a program controlling the electronic circuit, or the like). Hereinafter, a case will be further described in which each of the function configuration units of theinspection apparatus100 is achieved by the combination of hardware and software.

FIG.5 is a diagram illustrating acomputer1000 for achieving theinspection apparatus100. Thecomputer1000 is any computer. For example, thecomputer1000 is a personal computer (PC), a server machine, a tablet terminal, a smartphone, or the like. Thecomputer1000 may be a dedicated computer designed for achieving theinspection apparatus100, or may be a general-purpose computer.

Thecomputer1000 includes abus1020, aprocessor1040, amemory1060, astorage device1080, an input/output interface1100, and anetwork interface1120. Thebus1020 is a data transmission path through which theprocessor1040, thememory1060, thestorage device1080, the input/output interface1100, and thenetwork interface1120 transmit and receive data to and from one another. However, a method of connecting theprocessor1040 and the like with one another is not limited to bus connection. Theprocessor1040 is a processor such as a central processing unit (CPU), a graphics processing unit (GPU), or a field-programmable gate array (FPGA). Thememory1060 is a main storage apparatus achieved by using a random access memory (RAM) or the like. Thestorage device1080 is an auxiliary storage apparatus achieved by using a hard disk drive, a solid state drive (SSD), a memory card, a read only memory (ROM), or the like. However, thestorage device1080 may be configured by hardware, such as a RAM, similar to the hardware configuring the main storage apparatus.

The input/output interface1100 is an interface for connecting thecomputer1000 with an input/output device. Thenetwork interface1120 is an interface for connecting thecomputer1000 to a communication network. The communication network is, for example, a local area network (LAN) or a wide area network (WAN). A method by which thenetwork interface1120 connects to the communication network may be wireless connection, or may be wired connection.

Thestorage device1080 stores a program module for achieving the function configuration units of theinspection apparatus100. Theprocessor1040 achieves a function relevant to each program module, by reading each of the program modules into thememory1060 and executing the program module.

Flow of Processing

A flow of processing executed by theinspection apparatus100 will be described separately for the learning phase and the inference phase.FIG.6 is a flowchart illustrating a flow of processing executed in the inference phase by theinspection apparatus100 according to the example embodiment 1. Thetransformation unit120 acquires the first image data10 (S102). Thetransformation unit120 detects theinspection object90 from the first image data10 (S104). Thetransformation unit120 generates thesecond image data20 by applying image processing to thefirst image data10 in such a way that a view of theinspection object90 satisfies a predetermined reference (S106). Thedetection unit150 detects, by using an identification model, an abnormality of theinspection object90 included in the second image data20 (S108).

FIG.7 is a flowchart illustrating a flow of processing executed in the learning phase by theinspection apparatus100 according to the example embodiment 1. Note that, S102 to S106 are the same as inFIG.6. After S106, thelearning unit130 learns an identification model by using the second image data20 (S110).

Acquisition of First Image Data10: S102

Thetransformation unit120 acquires thefirst image data10. Herein, as thefirst image data10, image data of any type can be used. For example, as a type of thefirst image data10, a color image, a grayscale image, a black-and-white image, a depth image, a temperature image, and the like can be employed. Hereinafter, a case in which thefirst image data10 are a color image will be described as an example, but a type of thefirst image data10 is not limited to a color image. Further, description will be given with an upper left of thefirst image data10 as an origin, and with a horizontal direction and a vertical direction of thefirst image data10 as an x-axis and a y-axis, respectively.

Thefirst image data10 are generated by capturing an image of theinspection object90 with a camera. Note that, for a method of capturing an image of an inspection object by using a camera, any method can be employed. For example, an image of theinspection object90 is captured by a camera mounted on an unmanned aircraft such as a drone or a manned aircraft such as a helicopter. Besides the above, for example, an image of theinspection object90 may be captured from the ground by a person such as an operator.

Thetransformation unit120 acquires thefirst image data10 by various methods. For example, thetransformation unit120 acquires thefirst image data10 from the camera described above. Besides the above, for example, thetransformation unit120 may access a storage apparatus in which thefirst image data10 generated by a camera are stored, and thereby acquire thefirst image data10 from the storage apparatus.

Detection of Inspection Object90: S104

Thetransformation unit120 detects theinspection object90 from the first image data10 (S104). Herein, for a technique for detecting a particular object from an image, a variety of existing techniques can be used. For example, a template image or information indicating a feature value and the like of theinspection object90 to be detected from thefirst image data10 is stored, in advance, in a storage apparatus accessible from thetransformation unit120. Thetransformation unit120 detects theinspection object90 from thefirst image data10 by use of these pieces of information.

Generation of Second Image Data20: S106

Thetransformation unit120 generates thesecond image data20 by applying a geometric transform to thefirst image data10 in such a way that a view of theinspection object90 satisfies a predetermined reference (S106). For example, thetransformation unit120 generates thesecond image data20 according to a flow as follows. First, thetransformation unit120 computes a parameter representing a view of theinspection object90 on thefirst image data10. Hereinafter, the parameter will be referred to as a first parameter. Furthermore, thetransformation unit120 acquires a reference parameter. The reference parameter represents a predetermined reference for a view of theinspection object90. Thetransformation unit120 generates thesecond image data20 in which theinspection object90 having a view satisfying the predetermined reference is included, based on the first parameter computed for theinspection object90 on thefirst image data10, and the reference parameter.

For example, the first parameter can be computed by a following method. Thetransformation unit120 performs edge detection on an image area (hereinafter, an inspection object image) representing theinspection object90 on thefirst image data10, and generates edge information consisting of coordinate value information and the like of the detected edge. For edge detection, for example, a variety of approaches such as a Sobel filter, a Laplacian filter, a Canny filter, and the like can be used.

Furthermore, thetransformation unit120 detects a shape of theinspection object90 by using the edge information, and computes a value of the first parameter expressing the detected shape. For example, for detection of a shape, a Hough transform and the like can be used. Herein, the Hough transform was classically for detecting a line, but has been generalized to enable detection of a desired shape. Hereinafter, a case in which it is known in advance that a shape of an inspection object is a line will be described as an example.

An equation (1) below is an equation of the Hough transform for detecting a line.
[Mathematical 1]
ρ=x·cos θ+y·sin θ  (1)

In the equation (1), ρ and θ are parameters for defining a line, and respectively represent a length and an angle of a normal line drawn from an origin to any line. In other words, ρ and θ can be handled as the first parameter representing one line.

x and y are coordinate values of a pixel on an image. The equation (1) indicates any line passing through coordinate values for (x,y) on an image, and is equivalent to one sine curve on a (ρ,θ) plane (hereinafter, referred to as a Hough space). Note that, when a point on an image is determined, a sine curve is uniquely determined. Further, when a curve related to two points is overlaid on the Hough space, a point of intersection of two curves is related to a line on an image passing through the two points simultaneously.

Thetransformation unit120 computes the first parameter (ρ and θ described above) for a line representing theinspection object90, by performing the Hough transform using the edge information acquired by performing edge detection on the inspection object image. For example, thetransformation unit120 substitutes coordinate values (x,y) of all pixels forming an edge for the equation (1), and superimposes a plurality of curves on the Hough space. Then, thetransformation unit120 computes, based on overlapping of the superimposed curves, the first parameter (ρ and θ) for each of one or more lines detected from the inspection object image.

For example, thetransformation unit120 determines a pair of ρ and θ that maximizes the number of overlapping curves superimposed on the Hough space, and determines the pair as the first parameter. In this case, one line is detected from the inspection object image.

Besides the above, for example, thetransformation unit120 may determine a pair of θ and p that maximizes the number of overlapping curves superimposed on the Hough space, and may detect a shape having a parameter of a degree of divergence from at least one among the pair being equal to or less than a predetermined threshold value. In this way, for example, a line substantially parallel to a first detected line can be further detected. When theinspection object90 is a cable or the like having a certain degree of thickness, a boundary between theinspection object90 and a background is represented by two lines substantially parallel to each other (seeFIGS.2 and3). According to the method described above, a parameter representing each of two lines in such a relationship can be computed.

Herein, thetransformation unit120 in the learning phase may use teaching data representing a target to be processed. Hereinafter, a case in which the teaching data are an image will be described as an example, but the teaching data are not limited thereto.

FIG.8 illustrates one example of teaching data.Teaching data30 illustrated inFIG.8 are prepared from thefirst image data10, and include a background31, a power transmission line32, an arc mark33, and a snow-resistant ring34 illustrated with different patterns. Theteaching data30 may be prepared, for example, by painting areas such as the background31, the power transmission line32, the arc mark33, and the snow-resistant ring34 in different colors. Theteaching data30 are prepared, for example, manually by an operator or the like. Theteaching data30 clearly indicate a boundary between an inspection object and a background. Thus, thetransformation unit120 can detect a shape of theinspection object90 more easily than detecting a shape of theinspection object90 from thefirst image data10.

FIG.9 illustrates one example of different teaching data. Teaching data40 illustrated inFIG.9 are prepared from thefirst image data10, and consist of a point41, a point42, a point43, and a point44 each input by an operator with a pointing device or the like. Thetransformation unit120 may detect, as a shape of theinspection object90, for example, a line45 passing through the point41 and the point42 and a line46 passing through the point43 and the point44 by using the teaching data40. The teaching data40 have an advantageous effect of reducing operation load on an operator in preparation of teaching data, in comparison with theteaching data30.

Further, the teaching data may be, for example, an area specified by an operator tracing, with a finger or a device such as a pen, on an image displayed on a display device such as a touch panel.

Note that, thetransformation unit120 may search for the first parameter of theinspection object90 detected from the currently processedfirst image data10, by using the first parameter of theinspection object90 acquired from the previously processedfirst image data10. For example, it is assumed that thefirst image data10 are a video frame acquired from a video camera. Herein, a specific example will be described by using, as an example, a case in which a shape of theinspection object90 is a line.

As an example, it is assumed that a parameter (ρ_t, θ_t) of a line is to be computed for theinspection object90 detected from thefirst image data10 acquired at a time t. Further, it is assumed that a parameter (ρ_t-1, θ_t-1) of a line is already computed for theinspection object90 detected from thefirst image data10 acquired at a time t-1. Herein, when frequency of generation of thefirst image data10 is high enough, such as when thefirst image data10 are a video frame generated by a video camera, a view of theinspection object90 detected from thefirst image data10 acquired at the time t is thought almost the same as a view of theinspection object90 detected from thefirst image data10 acquired at the time t-1.

In view of the above, in this case, thetransformation unit120 searches for (ρ_t, θ_t) with (ρ_t-1, θ_t-1) as a reference. For example, by using predetermined threshold values α and β, thetransformation unit120 searches for a value for ρ_twithin a range of ρ_t-1±α and a value for θ_twithin a range of θ_t-1±β, respectively. According to the method, a search range (search space) for a parameter (ρ, θ) of a line can be limited, and thus, an advantageous effect of enabling improvement in processing speed can be acquired.

FIG.10 illustrates one example of a result of detecting a line on thefirst image data10. Herein, a line51 and a line52 are detected and are superimposed on thefirst image data10. Note that, in the present example, an angle53 formed by the detected line and a vertical direction of the image is θ, and a distance54 between the line51 and the line52 is d.

Herein, the distance d between two lines can be also used as the first parameter. Specifically, as will be described later, a ratio of enlargement or shrinkage to be applied to an inspection object image can be determined based on the distance d between two lines.

Thetransformation unit120 generates, by using the first parameter computed for theinspection object90 on thefirst image data10, thesecond image data20 in which theinspection object90 having a view satisfying a predetermined reference is included. For example, it is assumed that theinspection object90 is represented by a line. In this case, for example, thetransformation unit120 generates thesecond image data20 in such a way that a parameter of a line representing theinspection object90 on thesecond image data20 satisfies a predetermined reference.

Generation of thesecond image data20 is performed by using a reference parameter representing the predetermined reference described above. Thetransformation unit120 generates thesecond image data20 by applying a predetermined geometric transform to thefirst image data10, and thereby a degree of divergence between a parameter (hereinafter, a second parameter) representing theinspection object90 on thesecond image data20 and the reference parameter reduces.

For example, it is assumed that a parameter (ρ_r, θ_r) representing a line as a reference is determined as a reference parameter. Further, it is assumed that a parameter (ρ₁, θ₁) of a line is acquired as the first parameter representing theinspection object90 on thefirst image data10. In this case, thetransformation unit120 generates thesecond image data20 by performing a geometric transform on thefirst image data10 in such a way that a parameter (ρ₂, θ₂) of a line as the second parameter representing theinspection object90 on thesecond image data20 has a small degree of divergence from the reference parameter (ρ_r, θ_r). The geometric transform is, for example, parallel translation, rotation, scaling, and the like. Hereinafter, a method of computing a geometric transform parameter will be described by usingFIG.10.

For example, thetransformation unit120 may determine an area between the line51 and the line52 illustrated inFIG.10 as an inspection object area, select any point within the inspection object area as a first point, select any point on thefirst image data10 as a second point, and compute a parameter for parallel translation in such a way that the selected first point meets the selected second point. Such parallel translation is useful, for example, for preventing a part of theinspection object90 from not being included in thesecond image data20 due to rotation or the like.

Besides the above, for example, thetransformation unit120 computes a difference between a reference angle preliminarily determined for an angle formed by a line representing theinspection object90 and a vertical direction of the image and the angle53 illustrated inFIG.10, and determines the computed difference as a parameter for rotation. Besides the above, for example, thetransformation unit120 may compute a difference between θ_rincluded in the reference parameter of the line described above and θ included in a parameter of a line computed for theinspection object90 detected from thefirst image data10, and may determine the computed difference as a parameter for rotation.

Besides the above, for example, thetransformation unit120 computes a ratio between a reference distance preliminarily determined for two lines representing theinspection object90 and the distance54 illustrated inFIG.10, and determines the computed ratio as a parameter for scaling. Besides the above, for example, thetransformation unit120 may compute a ratio between ρ_rincluded in the reference parameter of the line described above and ρ computed for theinspection object90 detected from thefirst image data10, and may determine the computed ratio as a parameter for scaling.

A method of computing a geometric transform parameter is not limited to the method described above, and any method that can transform a view of theinspection object90 detected from thefirst image data10 into a view satisfying a predetermined reference can be used. For example, it is assumed that a parameter (ρ₁, θ₁) representing a shape of a line is acquired for theinspection object90 on thefirst image data10 and (ρ_r, θ_r) is determined as a reference parameter. In this case, for example, when a line determined by (ρ₁, θ₁) is transformed by an appropriate affine transformation into a line determined by (ρ_r, θ_r), a line representing theinspection object90 on thefirst image data10 can meet a line as a reference. In view of this, for example, thetransformation unit120 computes a transformation matrix for transforming (ρ₁, θ₁) into (ρ_r, θ_r) as a geometric transform parameter. Then, thetransformation unit120 generates thesecond image data20 by performing an affine transformation represented by the computed transformation matrix on thefirst image data10. Herein, for an approach to compute a transformation matrix for a geometric transform based on parameters of a line before and after the geometric transform, an existing approach can be used.

FIG.11 illustrates one example of thesecond image data20 according to the present example embodiment. In the present example, a parallel-translational geometric transform is performed in such a way that a center point of theinspection object90 meets a center point of the image. Further, in the present example, a rotational geometric transform is performed in such a way that an angle63 formed by a line61 and a vertical direction of the image becomes ∂′, with the center point of the image as an origin. Further, in the present example, a scaling geometric transform is performed in such a way that a distance64 between two detected lines becomes d′.

Herein, the geometric transform performed by thetransformation unit120 may be a part of the parallel translation, the rotation, and the scaling described above. For example, it is assumed that thetransformation unit120 performs parallel translation and rotation but does not perform scaling on thefirst image data10. In this case, since an image is not transformed by using a parameter for scaling, an inspection object has a size different for each image. In view of this, an analysis area as a unit for an analysis in learning or inference may have a size determined for each image.

The size of the analysis area is determined, for example, based on a size of theinspection object90 on an image. For example, it is assumed that the distance54 between the line51 and the line52 illustrated inFIG.10 is the size of the analysis area. Specifically, it is assumed that, when the distance54 is 32 pixels, the size of the analysis area is 32×32 pixels. However, the size of the analysis area can be any size suitable for detection of an abnormality, and is not limited to a size determined by the method described above.

Learning of Identification Model: S110

In the learning phase, thelearning unit130 learns an identification model by using the second image data20 (S110). The identification model extracts a plurality of local area images (hereinafter, patch images) from thesecond image data20, and identifies an abnormality and normality by using the extracted local area images. The patch image represents one analysis area. Herein, in order to learn the identification model, a label indicating whether the patch image is abnormal or normal (whether a part of theinspection object90 represented by the patch image is normal) is prepared for each patch image. The label is prepared, for example, manually by an operator or the like. Hereinafter, the label will be referred to as a teaching label. The teaching label may be referred to as correct answer data and the like.

In the following description, the teaching label indicating an abnormality will be represented as “1”, and the teaching label indicating normality will be represented as “0”. However, the teaching label is not limited to a case of binary representation as 0 or 1. For example, the teaching label may be a value (for example, any real number between 0 and 1 inclusive) representing a degree of normality or a degree of abnormality of a part of theinspection object90 represented by the patch image.

The identification model is learned by using an abnormal patch image and a normal patch image, and is achieved by any model such as a binary classifier or a multinomial classifier that can classify the patch images into two types or more. For example, the identification model is achieved by a convolutional neural network. Hereinafter, a case in which the identification model is achieved by a convolutional neural network will be described as an example.

In the following description, a result of determination as abnormal will be represented as “1”, and a result of determination as normal will be represented as “0”. However, similarly to the teaching label, a result of determination is also not limited to binary representation as normal or abnormal, and may represent a degree of normality or a degree of abnormality.

Thelearning unit130 inputs the patch image to the identification model. Consequently, a result of determination indicating whether the input patch image is normal or abnormal is output from the identification model. Herein, as described above, an analysis area may have a size different for eachsecond image data20. Thus, the patch image may have a size different for eachsecond image data20. In view of this, a reference size for the patch image may be determined in advance, and thelearning unit130 may enlarge or shrink the size of the patch image in such a way as to meet the reference size, and may thereafter input the patch image to the identification model.

Thelearning unit130 updates (learns) the identification model, based on the result of determination output from the identification model and the teaching label. In updating of the identification model, for example, a parameter set in the identification model is updated. Specifically, thelearning unit130 updates the identification model in such a way as to minimize an error representing a degree of divergence between the result of determination on the patch image and the teaching label related to the patch image. For example, when the result of determination on the patch image including an abnormality (the patch image with the related teaching label of “1”) is a value “1” representing abnormality, a degree of divergence between the result of determination and the teaching label is 0. On the other hand, when the result of determination on the patch image including an abnormality is a value “0” representing normality, a degree of divergence between the result of determination and the teaching label is 1.

Herein, for a method of defining the error for use in updating of the identification model, various methods can be employed in response to a type of the identification model. For example, a loss function E for computing an error between the result of determination and the teaching label is defined in advance. Thelearning unit130 optimizes, based on the result of determination and the teaching label, the parameter in the identification model in such a way as to minimize the loss function E. Herein, for a specific method of updating the parameter in the identification model by using a loss function, a variety of existing methods can be used.

For example, the loss function E can be defined by an equation (2) below.

$\begin{array}{cc}\left[\mathrm{Mathematical}2\right]& \\ E=-\sum _{n=1}^N\sum _{k^{\prime }=1}^K{t}_{k^{\prime }n}\log p_{k^{\prime }}\left(x_n\right)& \left(2\right)\end{array}$

In the equation (2), N represents the number of pieces of data, and K represents the total number of classes for classification. Further, t_k′nin the equation (2) is a vector of a correct answer label for n-th input data, and is a so-called 1-of-k vector (a vector with only a certain element being 1 and other elements being 0). A class having a value of 1 in t_k′nrepresents a correct answer class. For example, it is assumed that k=0 is a class representing normality and k=1 is a class representing abnormality, in a case of binary identification as abnormal or normal.

In this case, a correct answer vector related to input data representing normality is (1, 0), and a correct answer vector related to input data representing abnormality is (0, 1).

P_k′(x_n) in the equation (2) indicates a probability that the n-th input data belong to a class k′. For example, as described above, it is assumed that k=0 is a class representing normality and k=1 is a class representing abnormality. In this case, P₀(x_n) represents a probability that input data x_nbelong to the class representing normality, and P₁(x_n) represents a probability that the input data x_nbelong to the class representing abnormality. The probability P_k′(x_n) is computed by an equation (3) illustrated below.

$\begin{array}{cc}\left[\mathrm{Mathematical}3\right]& \\ P_k\left(x_n\right)=\frac{e^{f_k\left(x_n\right)}}{{\sum }_{k^{\prime }=1}^K{e}^{f_{k^{\prime }}\left(x_n\right)}}& \left(3\right)\end{array}$

In the equation (3), f_k(x_n) is an output value of a class k for the n-th input data x_n.

Detection of Abnormality (Execution of Inference): S108

In the inference phase, thedetection unit150 detects, by using thesecond image data20 and an identification model, an abnormal part of theinspection object90 included in the second image data20 (S108). Specifically, thedetection unit150 extracts a plurality of patch images from thesecond image data20 and inputs each of the patch images to an identification model, similarly to thelearning unit130. With this, a result of determination on each patch image is output from the identification model. For example, the result of determination represents whether the input patch image is normal or abnormal, or represents a degree of normality or a degree of abnormality of the input patch image. Herein, similarly to the learning, an analysis area may have a size different for eachsecond image data20. In view of this, in this case, for example, thedetection unit150 enlarges or shrinks the patch image acquired from thesecond image data20 in such a way as to meet a reference size, and inputs the patch image to the identification model.

Output of Result

Theinspection apparatus100 generates output information, based on a result of detection by thedetection unit150. Herein, a function configuration unit that generates and outputs output information will be referred to as an output unit.FIG.12 is a block diagram illustrating a function configuration of theinspection apparatus100 including anoutput unit160.

Theoutput unit160 generates output information, based on a result of determination on each patch image acquired from an identification model. For example, the output information indicates one or more pairs of “identification information of thesecond image data20 from which an abnormality is detected and a position of an abnormal part detected from the second image data20 (a position of the patch image from which the abnormality is detected)”. For the identification information of thesecond image data20, any information capable of identifying an image can be used. For example, the identification information of thesecond image data20 is a file name given to thesecond image data20. Besides the above, for example, the identification information of thesecond image data20 may be a file name or the like of thefirst image data10 being a generation source of thesecond image data20. Further, when thefirst image data10 are video frames constituting a video, a combination or the like of a file name of a video file including thefirst image data10 and a frame number of thefirst image data10 can be also used for identification information of thefirst image data10.

The output information may further indicate identification information of thesecond image data20 from which no abnormality is detected. Use of such output information enables recognition of whether an abnormality is detected from eachsecond image data20 and recognition of an abnormal part of thesecond image data20 from which an abnormality is detected.

Note that, instead of thesecond image data20, information on thefirst image data10 used for generation of thesecond image data20 may be output. For example, when an abnormality is detected from certainsecond image data20, thedetection unit150 generates output information indicating a pair of “identification information of thefirst image data10 being a transformation source of thesecond image data20 and a position on thefirst image data10 relevant to an abnormal part detected from thesecond image data20”. Herein, a position on thefirst image data10 relevant to an abnormal part detected from thesecond image data20 can be determined, for example, by performing, on the abnormal part detected from thesecond image data20, an inverse transform of a geometric transform performed to generate thesecond image data20 from thefirst image data10.

The output information is output by any output method. For example, the output information is stored in any storage apparatus. Besides the above, for example, the output information is output to a display apparatus connected with theinspection apparatus100. Besides the above, for example, the output information may be transmitted from theinspection apparatus100 to another apparatus. In this case, the output information is displayed on, for example, a display apparatus connected with the another apparatus.

FIG.13 is a diagram illustrating one example of output information. The output information inFIG.13 indicates information on each patch image determined as abnormal. Specifically, identification information of thesecond image data20 in which the patch image is included, a score (such as a probability that the patch image represents an abnormality) output by an identification model for the patch image, and a position of the patch image on thesecond image data20 are indicated. Note that, inFIG.13, the position of the patch image is represented by coordinates of four corners of the patch image. In this regard, information representing the position of the patch image is not limited to a method of representation by coordinates of four corners, but any method of determining an area on an image can be used. For example, the position of the patch image can be determined also by three values being coordinates of an upper-left corner, a width, and a height.

The output information is not limited to the above-described contents. For example, the output information may be information including thesecond image data20 or the like on which a display representing an abnormal part is superimposed.FIG.14 is a diagram illustrating thesecond image data20 on which a display representing an abnormal part is superimposed.FIG.14 includes anoutput screen70 being output to a display apparatus or the like connected with theinspection apparatus100. Theoutput screen70 includes adisplay area71 in which a part or whole of thesecond image data20 is displayed. Thedisplay area71 includes at least animage area72 that represents an abnormal part detected from thesecond image data20. Theimage area72 represents a patch image determined as abnormal among patch images acquired from thesecond image data20.

Herein, inFIG.14, theimage area72 is surrounded by a frame. The frame is equivalent to the “display representing an abnormal part” described above. Such a display enables a user of theinspection apparatus100 to easily recognize an abnormal part.

Further, on theoutput screen70, ascore73 is displayed near theimage area72. Thescore73 represents a score output by an identification model for the related image area72 (i.e., a patch image). Such a display of the score near the display representing an abnormal part enables a user of theinspection apparatus100 not only to recognize a part determined as abnormal, but also to recognize a degree of probability that the part is abnormal. With this, efficient measure for an abnormality of theinspection object90 can be performed such as, for example, preferentially checking of an abnormal part having a high score.

Theoutput screen70 further includes adisplay area74 in which metadata on thesecond image data20 from which an abnormality is detected are displayed. Specifically, inFIG.14, a file name, a frame number, a capturing date and time, and a capturing location are indicated. Herein, these pieces of information are information on thefirst image data10 being a transformation source of thesecond image data20.

Theoutput screen70 further includesswitch buttons76 and77 for switching thesecond image data20 displayed in thedisplay area71 and adelete button75. Thedelete button75 is a button for changing a result of determination on theimage area72 to “normal” when a user viewing theoutput screen70 determines that theimage area72 is misrecognized (i.e., theimage area72 does not represent an abnormal part). Specifically, a user clicks thedelete button75, by using a pointing device or the like, after specifying theimage area72 on which an erroneous result of determination is made. In this way, a result of determination on a patch image related to the specifiedimage area72 is changed from abnormal to normal. Note that, in this case, instead of overwriting a result of determination on a patch image indicated by output information, a result of determination (i.e., normal) by a user may be added to the output information while leaving the result of determination.

Herein, thelearning unit130 may learn by using a patch image for which a result of determination is changed by a user in this way. In other words, thelearning unit130 learns an identification model by use of learning data in which a teaching label as “normal” is related to theimage area72 where thedelete button75 is depressed by a user. In this way, the identification model has a lower probability of causing a similar determination error, and thus, accuracy in detection of an abnormality of theinspection object90 by using the identification model is improved.

Theoutput unit160 may determine a patch image to be displayed as theimage area72, based on a score related to the patch image. For example, theoutput unit160 displays, as theimage area72, patch images each having a score equal to or more than a predetermined threshold value. Besides the above, for example, theoutput unit160 may display, as theimage area72, patch images within a predetermined rank in descending order of related scores. The predetermined threshold value and the predetermined rank may be fixedly determined, or may be specifiable by a user of theinspection apparatus100.

Further, theoutput unit160 may change a type, a color, and the like of a display to be superimposed on a patch image in response to a score. For example, when a score represents a degree of abnormality, a domain of the score is divided into five numerical ranges, and different colors are allocated to the numerical ranges. For example, “blue”, “green”, “yellow”, “orange”, and “red” are allocated in ascending order of the degree of abnormality. Theoutput unit160 determines, for each patch image, a numerical range to which a score of the patch image belongs, and superimposes, on the patch image, a display of a color related to the determined numerical range.

Herein, it is assumed that, when a patch image is extracted from thesecond image data20, extraction is performed in such a way that adjacent patch images overlap one another. Then, it is assumed that patch images having a score of a predetermined threshold value or more overlap one another. In this case, theoutput unit160 may display, as theimage area72, only some (for example, one patch image having a maximum score) of the mutually overlapping patch images. Besides the above, for example, regarding patch images overlapping one another in this way, one image area represented by a sum of the mutually overlapping patch images may be displayed as theimage area72.

Besides the above, for example, theoutput unit160 may represent a degree of overlapping of the mutually overlapping patch images by using a heat map or the like.FIG.15 is one example of a degree of overlapping of patch images illustrated by a heat map. In an example of a heat map78 illustrated inFIG.15, as overlapping of patch images increases, density of black dots becomes larger.

Herein, in the present example embodiment, for convenience, the heat map is illustrated with density of black dots. However, theoutput unit160 may express the heat map with a gray scale or colors. For example, as a degree of overlapping of patch images increases, color changes in order of “blue”, “green”, “yellow, “orange”, and “red”.

Herein, it is assumed that a result of determination on a patch image is represented by a value between 0 to 1 in response to a degree of abnormality. In this case, theoutput unit160 computes, for example, a degree of overlapping of patch images by weighted addition of results of determination on the patch images. Specifically, when an amount of x-axis movement and an amount of y-axis movement of an analysis area are quarter a width and a height of the analysis area, respectively, sixteen times of inference are performed on any area of an inspection object, and thus, a total sum of results of determination may be divided by 16.

Other Functions

Theinspection apparatus100 may further include a function described below. For example, theinspection apparatus100 may include a function of accepting an input for checking that the first parameter of theinspection object90 has been correctly detected from thefirst image data10. A function configuration unit having the function will be referred to as a checking unit.FIG.16 is a block diagram illustrating a function configuration of theinspection apparatus100 including achecking unit170.

FIG.17 is a diagram illustrating a screen output by thechecking unit170. Ascreen80 illustrated inFIG.17 includes adisplay area81 that displays an image. The image displayed in thedisplay area81 is, for example, an image in which a display representing a shape of theinspection object90 is superimposed on thefirst image data10. In an example illustrated inFIG.17, a plurality of lines (aline82 and a line83) detected from thefirst image data10 are superimposed on thefirst image data10, as a display representing a shape of theinspection object90.

Further, thescreen80 may include aninformation display area85 that displays information (a file name and the like) relating to thefirst image data10. Thescreen80 may further include, for example, a shapeinformation display area84 that displays a detail of a shape of theinspection object90 displayed in thedisplay area81. Specifically, when a shape of theinspection object90 is theline82 and theline83 illustrated inFIG.17, thescreen80 may display, in the shapeinformation display area84, for example, coordinate values or the like of a starting point and an end point on the image. In addition, thescreen80 may include, for example, abutton86 that accepts that the image displayed in thedisplay area81 has been checked by an operator.

For example, when a shape displayed in thedisplay area81 is not superimposed on an appropriate position of the inspection object90 (i.e., when the first parameter is not appropriate), an operator may select and move the shape to an appropriate position by using a pointing device or the like. Further, an operator may correct a detail of a shape of an inspection object displayed in the shapeinformation display area84 by using an input device such as a keyboard.

Use of thescreen80 enables checking that the first parameter has been appropriately computed from theinspection object90. Further, when the computed first parameter is incorrect, the computed first parameter can be corrected to a correct parameter. With this, a probability that thesecond image data20 are generated correctly, that is, a probability that a view of theinspection object90 is geometrically transformed correctly in such a way as to satisfy a predetermined reference can be increased. Thus, an abnormality of theinspection object90 can be detected with higher accuracy.

When theinspection apparatus100 includes thechecking unit170, thechecking unit170 operates before thesecond image data20 are generated after the first parameter is computed by thetransformation unit120. In other words, thesecond image data20 are generated after the above-described checking is performed by an operator or the like.

While the example embodiments of the present invention have been described with reference to the drawings, the above-described example embodiments are exemplified of the present invention, and a combination of the above-described example embodiments or various configurations other than the above may be employed.

The whole or part of the above-described example embodiments can be described as, but not limited to, the following supplementary notes.

1. An inspection apparatus including:

a transformation unit that detects an inspection object from first image data in which the inspection object is included, and generates second image data by performing a geometric transform on the first image data in such a way that a view of the detected inspection object becomes a view satisfying a predetermined reference; and

a detection unit that detects, by using an identification model for detecting an abnormality of the inspection object, an abnormality of the inspection object included in the second image data, wherein

the identification model is learned by using image data in which an inspection object having a view satisfying a predetermined reference is included.

2. An inspection apparatus including:

a transformation unit that detects an inspection object from first image data in which the inspection object is included, and generates second image data by performing a geometric transform on the first image data in such a way that a view of the detected inspection object becomes a view satisfying a predetermined reference; and

a learning unit that learns, by using the second image data, an identification model for detecting an abnormality of the inspection object.

3. The inspection apparatus according to supplementary note 1 or 2, wherein

the transformation unit

• • computes a first parameter representing a view of the inspection object on the first image data,
  • acquires a reference parameter representing the predetermined reference, and
  • determines, based on the first parameter and the reference parameter, a geometric transform to be applied to the first image data in such a way that a degree of divergence between a second parameter representing a view of the inspection object on the second image data and the reference parameter is reduced.
    4. The inspection apparatus according to supplementary note 3, wherein

the first parameter includes any one of a position where a line representing the inspection object and any one edge of the first image data intersect each other, a size of an angle formed by a line representing the inspection object and any one edge of the first image data, a length of a normal line drawn from an origin to a line representing the inspection object, and an angle of the normal line.

5. The inspection apparatus according to any one of supplementary notes 1 to 4, wherein

the transformation unit determines, by using teaching data in which an image area representing the inspection object on the first image data is specified, a geometric transform for generating the second image data from the first image data.

6. The inspection apparatus according to any one of supplementary notes 1 to 5, further including

a checking unit that outputs a display indicating an image area representing the inspection object detected from the first image data, and accepts an input for correcting an image area representing the inspection object.

7. The inspection apparatus according to any one of supplementary notes 1 to 6, further including

an output unit that determines an area on the first image data equivalent to an abnormal part of the inspection object detected from the second image data, and outputs the first image data in which a predetermined display is superimposed on the determined area.

8. A control method executed by a computer, including:

a transformation step of detecting an inspection object from first image data in which the inspection object is included, and generating second image data by performing a geometric transform on the first image data in such a way that a view of the detected inspection object becomes a view satisfying a predetermined reference; and

a detection step of detecting, by using an identification model for detecting an abnormality of the inspection object, an abnormality of the inspection object included in the second image data, wherein

the identification model is learned by using image data in which an inspection object having a view satisfying a predetermined reference is included.

9. A control method executed by a computer, including:

a transformation step of detecting an inspection object from first image data in which the inspection object is included, and generating second image data by performing a geometric transform on the first image data in such a way that a view of the detected inspection object becomes a view satisfying a predetermined reference; and

a learning step of learning, by using the second image data, an identification model for detecting an abnormality of the inspection object.

10. The control method according to supplementary note 8 or 9, further including:

in the transformation step,

• • computing a first parameter representing a view of the inspection object on the first image data;
  • acquiring a reference parameter representing the predetermined reference; and
  • determining, based on the first parameter and the reference parameter, a geometric transform to be applied to the first image data in such a way that a degree of divergence between a second parameter representing a view of the inspection object on the second image data and the reference parameter is reduced.
    11. The control method according tosupplementary note 10, wherein

the first parameter includes any one of a position where a line representing the inspection object and any one edge of the first image data intersect each other, a size of an angle formed by a line representing the inspection object and any one edge of the first image data, a length of a normal line drawn from an origin to a line representing the inspection object, and an angle of the normal line.

12. The control method according to any one of supplementary notes 8 to 11, further including,

in the transformation step, determining, by using teaching data in which an image area representing the inspection object on the first image data is specified, a geometric transform for generating the second image data from the first image data.

13. The control method according to any one of supplementary notes 8 to 12, further including

a checking step of outputting a display indicating an image area representing the inspection object detected from the first image data, and accepting an input for correcting an image area representing the inspection object.

14. The control method according to any one of supplementary notes 8 to 13, further including

an output step of determining an area on the first image data equivalent to an abnormal part of the inspection object detected from the second image data, and outputting the first image data in which a predetermined display is superimposed on the determined area.

15. A program causing a computer to execute each of steps in the control method according to any one of supplementary notes 8 to 14.

Claims (20)

What is claimed is:

1. An inspection apparatus comprising:

a memory storing instructions; and

a processor configured to execute the instructions to:

detect an inspection object from first image data in which the inspection object is included;

generate second image data by performing a geometric transform on the first image data in such a way that a view of the detected inspection object satisfies a predetermined reference; and

detect, by using an identification model, an abnormality of the inspection object included in the second image data, wherein

the identification model is learned by using image data in which learning inspection objects having views satisfying the predetermined reference are included, and

in the geometric transform, a degree of divergence between a second parameter representing a second view of the inspection object on the second image data and the reference becomes smaller than a degree of divergence between a first parameter representing a first view of the inspection object on the first image data and the reference.

2. The inspection apparatus according toclaim 1, wherein

the second image data is generated by

computing the first parameter representing the first view of the inspection object on the first image data;

acquiring a reference parameter representing the predetermined reference; and

determining, based on the first parameter and the reference parameter, the geometric transform to apply to the first image data in such a way that the degree of divergence between the second parameter representing the second view of the inspection object on the second image data and the reference parameter is reduced.

3. The inspection apparatus according toclaim 2, wherein

the first parameter includes any one of a position where a line representing the inspection object and any one edge of the first image data intersect each other, a size of an angle formed by the line representing the inspection object and the any one edge of the first image data, a length of a normal line drawn from an origin to the line representing the inspection object, and an angle of the normal line.

4. The inspection apparatus according toclaim 1, wherein

the inspection object is detected from the first image data by using teaching data in which an image area representing the inspection object on the first image data is specified.

5. The inspection apparatus according toclaim 1, wherein the processor is configured to execute the instructions to further:

display an image area representing the inspection object detected from the first image data; and

accept an input for correcting the image area representing the inspection object.

6. The inspection apparatus according toclaim 1, wherein the processor is configured to execute the instructions to further:

determine an area on the first image data equivalent to an abnormal part of the inspection object detected from the second image data and

display a predetermined display superimposed on the determined area.

7. The inspection apparatus according toclaim 1, wherein

the geometric transform uses the second parameter as the reference.

8. An inspection apparatus comprising:

a memory configured to store instructions; and

a processor configured to execute the instructions including:

detect an inspection object from first image data in which the inspection object is included;

generate second image data by performing a geometric transform on the first image data in such a way that a view of the detected inspection object satisfies a predetermined reference; and

learn, by using the second image data, an identification model for detecting an abnormality of the inspection object, wherein

in the geometric transform, a degree of divergence between a second parameter representing a second view of the inspection object on the second image data and the reference becomes smaller than a degree of divergence between a first parameter representing a first view of the inspection object on the first image data and the reference.

9. The inspection apparatus according toclaim 8, wherein

the second image data is generated by

computing the first parameter representing the first view of the inspection object on the first image data;

acquiring a reference parameter representing the predetermined reference; and

determining, based on the first parameter and the reference parameter, the geometric transform to be applied to apply to the first image data in such a way that the degree of divergence between the second parameter representing the second view of the inspection object on the second image data and the reference parameter is reduced.

10. The inspection apparatus according toclaim 9, wherein

the first parameter includes any one of a position where a line representing the inspection object and any one edge of the first image data intersect each other, a size of an angle formed by the line representing the inspection object and the any one edge of the first image data, a length of a normal line drawn from an origin to the line representing the inspection object, and an angle of the normal line.

11. The inspection apparatus according toclaim 8, wherein

the inspection object is detected from the first image data by using teaching data in which an image area representing the inspection object on the first image data is specified.

12. The inspection apparatus according toclaim 8, wherein the processor is configured to execute the instructions to further:

display indicating an image area representing the inspection object detected from the first image data; and

accept an input for correcting the image area representing the inspection object.

13. The inspection apparatus according toclaim 8, wherein the processor is configured to execute the instructions to further:

determine an area on the first image data equivalent to an abnormal part of the inspection object detected from the second image data and

display a predetermined display superimposed on the determined area.

14. A control method comprising:

detecting, by a processor, an inspection object from first image data in which the inspection object is included;

generating, by the processor, second image data by performing a geometric transform on the first image data in such a way that a view of the detected inspection object satisfies a predetermined reference; and

detecting, by the processor using an identification model, an abnormality of the inspection object included in the second image data, wherein

the identification model is learned by using image data in which learning inspection objects having views satisfying the predetermined reference are included, and

in the geometric transform, a degree of divergence between a second parameter representing a second view of the inspection object on the second image data and the reference becomes smaller than a degree of divergence between a first parameter representing a first view of the inspection object on the first image data and the reference.

15. The control method according toclaim 14, wherein further comprising:

generating the second image data comprises:

computing the first parameter representing the first view of the inspection object on the first image data;

acquiring a reference parameter representing the predetermined reference; and

determining, based on the first parameter and the reference parameter, the geometric transform to apply to the first image data in such a way that the degree of divergence between the second parameter representing the second view of the inspection object on the second image data and the reference parameter is reduced.

16. The control method according toclaim 15, wherein

the first parameter includes any one of a position where a line representing the inspection object and any one edge of the first image data intersect each other, a size of an angle formed by the line representing the inspection object and the any one edge of the first image data, a length of a normal line drawn from an origin to the line representing the inspection object, and an angle of the normal line.

17. The control method according toclaim 14, wherein

the inspection object is detected from the first image data by using teaching data in which an image area representing the inspection object on the first image data is specified.

18. The control method according toclaim 14, further comprising:

displaying, by the processor, an image area representing the inspection object detected from the first image data; and

accepting, by the processor, an input for correcting the image area representing the inspection object.

19. The control method according toclaim 14, further comprising:

determining, by the processor, an area on the first image data equivalent to an abnormal part of the inspection object detected from the second image data; and

displaying, by the processor, a predetermined display superimposed on the determined area.

20. A non-transitory computer readable medium storing a program executable by a computer to perform the control method according toclaim 14.

Images